In his very informative post, Recommendation Systems for Large Datasets @ravishah1 explains how re-ranking models are the industry standard for dealing with datasets like we are presented with in this competition, that is ones with high cardinality categories!
\nThis starter pack builds exactly such a model!
\nI preprocess the data using polars (mostly to try out the library and to decrease memory footprint vs pandas, the added speed is also nice 🙂).
I then proceed to train an LGBMRanker, output predictions and submit to the competition.
The notebook provides all the low-level plumbing that you can build on 🙂 The next steps here might include using co-visitation matrices for candidate generation or scoring. One can also generate additional features (such as times an item was clicked or put in cart, popularity scores, temporal features, etc).
\nHappy hacking! 🙂
\nIn his very informative post, Recommendation Systems for Large Datasets @ravishah1 explains how re-ranking models are the industry standard for dealing with datasets like we are presented with in this competition, that is ones with high cardinality categories!
\nThis starter pack builds exactly such a model!
\nI preprocess the data using polars (mostly to try out the library and to decrease memory footprint vs pandas, the added speed is also nice 🙂).
I then proceed to train an LGBMRanker, output predictions and submit to the competition.
The notebook provides all the low-level plumbing that you can build on 🙂 The next steps here might include using co-visitation matrices for candidate generation or scoring. One can also generate additional features (such as times an item was clicked or put in cart, popularity scores, temporal features, etc).
\nHappy hacking! 🙂
\nThanks for the starter Radek. In the last Kaggle recsys comp, ML reranker beat heuristic reranker. So i think the winning teams will be using an approach like this.
", "rawMarkdown": "Thanks for the starter Radek. In the last Kaggle recsys comp, ML reranker beat heuristic reranker. So i think the winning teams will be using an approach like this.", "votes": null }, { "id": "2032933", "postDate": "11/16/2022 22:27:29", "content": "Thank you, @cdeotte! 🙌 That is a very valuable piece of information 🙂
", "rawMarkdown": "Thank you, @cdeotte! 🙌 That is a very valuable piece of information 🙂", "votes": null } ], "comments": [ { "id": 2032531, "author_name": "cdeotte", "author_url": "", "post_date": "11/16/2022 16:52:35", "content": "Thanks for the starter Radek. In the last Kaggle recsys comp, ML reranker beat heuristic reranker. So i think the winning teams will be using an approach like this.
", "votes": null, "replies": [ { "id": 2032933, "author_name": "radek1", "author_url": "", "post_date": "11/16/2022 22:27:29", "content": "Thank you, @cdeotte! 🙌 That is a very valuable piece of information 🙂
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2030220": "In his very informative post, [Recommendation Systems for Large Datasets](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364721) @ravishah1 explains how re-ranking models are the industry standard for dealing with datasets like we are presented with in this competition, that is ones with high cardinality categories!\n\n[This starter pack](https://www.kaggle.com/code/radek1/polars-proof-of-concept-lgbm-ranker/notebook) builds exactly such a model!\n\nI preprocess the data using [polars](https://www.pola.rs/) (mostly to try out the library and to decrease memory footprint vs `pandas`, the added speed is also nice 🙂).\n\nI then proceed to train an `LGBMRanker`, output predictions and submit to the competition.\n\n[The notebook](https://www.kaggle.com/code/radek1/polars-proof-of-concept-lgbm-ranker/notebook) provides all the low-level plumbing that you can build on 🙂 The next steps here might include using co-visitation matrices for candidate generation or scoring. One can also generate additional features (such as times an item was clicked or put in cart, popularity scores, temporal features, etc).\n\nHappy hacking! 🙂 \n\n### Other resources you might find useful:\n\n* [💡 [2 methods] How-to ensemble predictions 🏅🏅🏅](https://www.kaggle.com/code/radek1/2-methods-how-to-ensemble-predictions)\n* [local validation tracks public LB perfecty -- here is the setup](https://www.kaggle.com/competitions/otto-recommender-system/discussion/364991)\n* [💡 For my friends from Twitter and LinkedIn -- here is how to dive into this competition 🐳](https://www.kaggle.com/competitions/otto-recommender-system/discussion/368560)\n* [Full dataset processed to CSV/parquet files with optimized memory footprint](https://www.kaggle.com/competitions/otto-recommender-system/discussion/363843)\n* [co-visitation matrix - simplified, imprvd logic 🔥](https://www.kaggle.com/code/radek1/co-visitation-matrix-simplified-imprvd-logic)\n* [💡 Word2Vec How-to [training and submission]🚀🚀🚀](https://www.kaggle.com/code/radek1/word2vec-how-to-training-and-submission)", "2032531": "Thanks for the starter Radek. In the last Kaggle recsys comp, ML reranker beat heuristic reranker. So i think the winning teams will be using an approach like this.", "2032933": "Thank you, @cdeotte! 🙌 That is a very valuable piece of information 🙂" }, "source": "meta" }